Collusion by mistake: Does algorithmic sophistication drive supra-competitive profits?
研究发现，自学习算法在重复互动中可能因探索不足而看似合谋，但更复杂的算法通过更充分的探索反而促进竞争，甚至利用看似合谋的算法。
A burgeoning literature shows that self-learning algorithms may, under some conditions, reach seemingly-collusive outcomes: after repeated interaction, competing algorithms earn supra-competitive profits, at the expense of efficiency and consumer welfare. This paper offers evidence that such behavior can stem from insufficient exploration during the learning process and that algorithmic sophistication might increase competition. In particular, we show that allowing for more thorough exploration does lead otherwise seemingly-collusive Q-learning algorithms to play more competitively. We first provide a theoretical illustration of this phenomenon by analyzing the competition between two stylized Q-learning algorithms in a Prisoner’s Dilemma framework. Second, via simulations, we show that some more sophisticated algorithms exploit the seemingly-collusive ones. Following these results, we argue that the advancement of algorithms in sophistication and computational capabilities may, in some situations, provide a solution to the challenge of algorithmic seeming collusion, rather than exacerbate it.